Recent progress in artificial intelligence has encouraged numerous attempts to understand and decode human visual system from brain signals. These prior works typically align neural activity independently with semantic and perceptual features extracted from images using pre-trained vision models. However, they fail to account for two key challenges: (1) the modality gap arising from the natural difference in the information level of representation between brain signals and images, and (2) the fact that semantic and perceptual features are highly entangled within neural activity. To address these issues, we utilize hyperbolic space, which is well-suited for considering differences in the amount of information and has the geometric property that geodesics between two points naturally bend toward the origin, where the representational capacity is lower. Leveraging these properties, we propose a novel framework, Hyperbolic Feature Interpolation (HyFI), which interpolates between semantic and perceptual visual features along hyperbolic geodesics. This enables both the fusion and compression of perceptual and semantic information, effectively reflecting the limited expressiveness of brain signals and the entangled nature of these features. As a result, it facilitates better alignment between brain and visual features. We demonstrate that HyFI achieves state-of-the-art performance in zero-shot brain-to-image retrieval, outperforming prior methods with Top-1 accuracy improvements of up to +17.3% on THINGS-EEG and +9.1% on THINGS-MEG.

Vision-language models have achieved remarkable performance across various tasks by leveraging large-scale multimodal training data. However, their ability to generalize to out-of-distribution (OOD) domains requiring expert-level knowledge remains an open challenge. To address this, we investigate cross-domain transfer learning approaches for efficiently adapting diffusion classifiers to new target domains demanding expert-level domain knowledge. Specifically, we propose ExpertDiff, a head-less model reprogramming technique that optimizes the instruction-following abilities of text-to-image diffusion models via learnable prompts, while leveraging the diffusion classifier objective as a modular plug-and-play adaptor. Our approach eliminates the need for conventional output mapping layers (e.g., linear probes), enabling seamless integration with off-the-shelf diffusion frameworks like Stable Diffusion. We demonstrate the effectiveness of ExpertDiff on the various OOD datasets (i.e., medical and satellite imagery). Furthermore, we qualitatively showcase ExpertDiff’s ability to faithfully reconstruct input images, highlighting its potential for both downstream discriminative and upstream generative tasks. Our work paves the way for effectively repurposing powerful foundation models for novel OOD applications requiring domain expertise.